I've only touched the surface of ML after a career change from translation to SE. But even with that superficial knowledge, I find it very cute when people start looking at announcements like GNMT as though translators are going to be completely displaced. If all you need is a "good enough" communication, then fine, but the second you have to start explicitly considering things like "what are the motivations of the person speaking" in order to come up with a good translation, it's pretty obvious how ill-equipped the ML toolset is for solving translation "completely."